Fix compatibility with latest torch nightly (#3666)

* remove ddp UT since ddp support is deprecated

* change torch-ccl branch to torch_ccl_dev_2.8

* update torch-ccl version

* Update dependency_version.json

* remove unused subprocess

* fix format

* fix ddp UT

---------

Co-authored-by: DiweiSun <105627594+DiweiSun@users.noreply.github.com>

Author: chunyuan-w

dependency_version.json

@@ -7,17 +7,17 @@
     "commit": "llvmorg-16.0.6"
   },
   "pytorch": {
-    "version": "2.8.0.dev20250323+cpu"
+    "version": "2.8.0.dev20250424+cpu"
   },
   "torchaudio": {
-    "version": "2.6.0.dev20250323+cpu"
+    "version": "2.6.0.dev20250424+cpu"
   },
   "torchvision": {
-    "version": "0.22.0.dev20250322+cpu"
+    "version": "0.22.0.dev20250423+cpu"
   },
   "torch-ccl": {
-    "commit": "torch_ccl_dev_2.6",
-    "version": "2.6.0+cpu"
+    "commit": "torch_ccl_dev_2.8",
+    "version": "2.8.0+cpu"
   },
   "deepspeed": {
     "version": "",
@@ -29,4 +29,4 @@
   "lm-evaluation-harness": {
     "commit": "cc9778fbe4fa1a709be2abed9deb6180fd40e7e2"
   }
-}
+}

tests/cpu/ipex-optimize-ddp-static-graph.py

@@ -54,4 +54,4 @@ def env2int(env_list, default=-1):
     loss.backward()
     opt.step()
     if i == 9:
-        print(f"Resume training successfully, final lose = {loss.item()}")
+        print(f"Resume training successfully, final lose = {loss.detach().item()}")